The iCub robot, which looks and learns like a young child, was presented with a new way to demonstrate its impressive learning skills – by giving it a weapon and teaching it how to shoot it. Don’t worry – it’s just a bow and arrows, a toy one at that, but it is an interesting look into the method such a robot goes through when testing out different solutions to get to the answer of a problem.
The new algorithm it was taught, ARCHER (Augmented Reward Chained Regression), gives it a specific goal in mind, such as hitting the bullseye of a target, and makes it use each previous experience of shooting the arrow to determine what is the best angle it can aim at to achieve maximum success. Humans also try to use a similar method when practicing with arrows, but the problem is our memories aren’t nearly as organized as a robot’s, and we can even struggle to recreate the shot we just fired. The iCub doesn’t have that problem, however, and so with this algorithm it steadily improves its performance and gets closer and closer to hitting the mark.
ARCHER controls the positioning and movements of the robot’s hands, though an inverse kinematics controller is needed for the arms. The iCub receives the results of its last attempt through its image processing technology, which uses the Gaussian Mixture Model to decipher different colors and shapes. The image is then sent back to the algorithm, which adjusts the robot to a new angle for the next try.
In the video you can see all this theory in action, where iCub, dressed in a funny Native American costume, shoots arrows at a bullseye 3.5 meters away. It starts out hitting the adjacent windows, but by the end learns how to hit the target dead-center.